Predicting adenine base editing efficiencies in different cellular contexts by deep learning

Abstract

Background: Adenine base editors (ABEs) enable the conversion of A•T to G•C base pairs. Since the sequence of the target locus influences base editing efficiency, efforts have been made to develop computational models that can predict base editing outcomes based on the targeted sequence. However, these models were trained on base editing datasets generated in cell lines and their predictive power for base editing in primary cells in vivo remains uncertain.

Results: In this study, we conduct base editing screens using SpRY-ABEmax and SpRY-ABE8e to target 2,195 pathogenic mutations with a total of 12,000 guide RNAs in cell lines and in the murine liver. We observe strong correlations between in vitro datasets generated by ABE-mRNA electroporation into HEK293T cells and in vivo datasets generated by adeno-associated virus (AAV)- or lipid nanoparticle (LNP)-mediated nucleoside-modified mRNA delivery (Spearman R = 0.83-0.92). We subsequently develop BEDICT2.0, a deep learning model that predicts adenine base editing efficiencies with high accuracy in cell lines (R = 0.60-0.94) and in the liver (R = 0.62-0.81).

Conclusions: In conclusion, our work confirms that adenine base editing holds considerable potential for correcting a large fraction of pathogenic mutations. We also provide BEDICT2.0 - a robust computational model that helps identify sgRNA-ABE combinations capable of achieving high on-target editing with minimal bystander effects in both in vitro and in vivo settings.

Keywords: CRISPR-Cas9 genome editing; Genomics; In vivo; Machine learning; Mouse.

Fig. 1

High-throughput ABE screening in HEK293T cells using target-matched sgRNA libraries. a Strategy of correcting pathogenic mutations without bystander editing by sgRNA tiling. The sgRNA not including the coding bystander within the editing window is shadowed darker. b Schematics of the ABE screen in HEK293T cells using plasmid transfection for ABE delivery. c Total editing efficiencies for each PAM in HEK293T cells after 10 days ABE selection with SpRY-ABE8e and SpRY-ABEmax (top row), SpG-ABE8e and SpG-ABEmax (middle row), SpCas9-ABE8e and SpCas9-ABEmax (bottom row). Y-axis indicates the 1st nucleotide of the PAM motif, the x-axis the 2nd and 3.^rd nucleotide of the PAM. d Editing window for SpRY-ABE8e and SpRY-ABEmax (top row), SpG-ABE8e and SpG-ABEmax (middle row), SpCas9-ABE8e and SpCas9-ABEmax (bottom row). Datasets were filtered for best PAMs (NRN for SpRY, NGN for SpG, and NGG for SpCas9). e Correction of pathogenic mutations in the library with- or without inducing non-silent bystander mutations for different base editors. Cut-offs were ≥ 10% for on-target editing and ≤ 0.5% for bystander editing. Target sites with on-target editing below 10% were defined as not corrected. Number of target sites (n) for SpRY-ABE8e: 11838, SpRY-ABEmax: 11497, SpG-ABE8e: 10287, SpG-ABEmax: 9400, SpCas9-ABE8e: 7540, SpCas9-ABEmax: 9702, ABE combined: 12000

Fig. 2

High-throughput ABE screening in the liver cells with target-matched sgRNA libraries reveals correlation to cell culture. a The sgRNA library was injected in p1 pups prior to ABE injection in juvenile mice. Editing rates were analysed by HTS. b Correlation of total A-to-G editing between the mRNA-LNP and AAV dataset with SpRY-ABE8e (n = 2176) and SpRY-ABEmax (n = 7247). The red line represents linear regression. c Violin plot of total editing efficiency for SpRY-ABE8e and SpRY-ABEmax in the indicated datasets. Datasets were filtered for most efficient PAMs (NRN) and mean editing efficiency is plotted (grey line). n for SpRY-ABE8e = 7882, 1623, 3459 and SpRY-ABEmax = 7644, 5170, 5852. d Total editing efficiency for each PAM present in the library for SpRY-ABE8e (left) and SpRY-ABEmax (right) for the mRNA-LNP and AAV datasets. e Editing window in the mRNA-LNP and AAV datasets are for SpRY-ABE8e (left) and SpRY-ABEmax (right) filtered for best PAMs (NRN). f Proportion of the different tri-nucleotide motifs for loci above mean editing efficiency (top) and below mean editing efficiency (bottom) for SpRY-ABE8e (left) and SpRY-ABEmax (right) of various screening methods

Fig. 3

Correlation of editing efficiencies between in vitro and in vivo ABE screening datasets. a Correlation of total A-to-G editing efficiency between in vivo (mRNA-LNP and AAV) and in vitro (HEK-Plasmid) screening datasets for SpRY-ABE8e (left, n = 2418, 5233) and SpRY-ABEmax (right, n = 7817, 8770). b Violin plots of total editing efficiency in mRNA-ABE datasets with SpRY-ABE8e (top) and SpRY-ABEmax (bottom) with 0.2 pmol, 1 pmol or 5 pmol mRNA transfection. Datasets were filtered for best PAMs (NRN) and mean editing efficiency is given (grey line). n for SpRY-ABE8e = 6361, 6424, 5730 and SpRY-ABEmax = 6159, 6322, 5961. c Correlation of total A-to-G editing efficiency between in vivo (mRNA-LNP and AAV) and in vitro (HEK-mRNA) screening datasets for SpRY-ABE8e (left, n = 2388, 5018) and SpRY-ABEmax (right, n = 7308, 7897). The red line in all plots represents linear regression

Fig. 4

Establishment and evaluation of BEDICT2.0, a machine learning model predicting ABE activity in vitro and in vivo. a Schematics of BEDICT2.0 machine learning algorithm. BEDICT2.0 includes an Efficiency Model (predicts total editing efficiency) and a Proportion Model (predicts distribution within the edited reads). Outputs of both models are combined to predict editing efficiency. b Comparison of the performance of BEDICT1.2 or BEDICT2.0 on various HEK-Plasmid test datasets generated in this study. c Comparison of the performance of BEDICT2.0 trained on either HEK-Plasmid and tested on the in vivo datasets, trained on HEK-mRNA and tested on the in vivo datasets or trained and tested on the in vivo datasets. d Editing efficiency predicted by BEDICT2.0 plotted against the measured efficiency for SpRY-ABE8e (top) and SpRY-ABEmax (bottom) for HEK-mRNA (5 pmol), mRNA-LNP or AAV datasets. The red line represents linear regression. e Comparison of BEDICT2.0 to other base editing prediction models on adenine base editing datasets from target-matched sgRNA library screens. Datasets used for comparison are SpCas9-ABEmax (mES-12kChar) [23] and SpCas9-ABE7.10 (HT-ABE Train) [24]. ML-models used for predicting ABE editing outcome: DeepABE [24], BE-Hive-ABE-HEK293T [23] and BEDICT2.0 (this study). f Total A-to-G editing efficiency at endogenous loci in various datasets correlated to BEDICT2.0 (trained on the HEK-plasmid dataset) predictions. n for Marquart-HEK293T [25]: 18, Song-HEK293T: 72, Song-U2OS: 22, Song-HCT116: 41 [24]. g Spearman and Pearson correlation of measured and predicted editing efficiencies with BE-HIVE [23], DeepABE [24] and BEDICT2.0 (trained on HEK-plasmid) on various datasets generated on endogenous loci. h Spearman and Pearson correlation of measured and predicted editing efficiencies of BE-HIVE [23] and DeepABE [24] on the different SpRY-ABEmax datasets. Datasets were filtered for protospacers with NGG PAMs for DeepABE, as the model can only be applied for NGG PAMs